GPT Coding Model Costs (2025-2026): Codex vs GPT-5.2/5.4/5.5 Pricing and Context

May 05, 2026

When you run LLM-assisted coding in production, you stop asking “Which model is best?” and start asking “Which model is best per dollar for my workload?” The tricky part is that coding agents burn tokens differently than chat: they ingest a lot of context, but they can also emit a lot of output (patches, tests, logs, explanations).

This post compares capability milestones and, more importantly, token pricing for OpenAI’s coding-focused models through May 5, 2026:

Codex models (gpt-5-codex … gpt-5.3-codex)
Flagship GPT models often used for coding (gpt-5.2, gpt-5.4, gpt-5.5)

Most of this article is about API token cost (input, cached input, output). At the end, there’s a short section on ChatGPT Plus, because subscription access is a different pricing model and it’s easy to mix them up.

Token Pricing Basics (The Only 3 Numbers That Matter)

OpenAI bills Codex model usage by token category:

Input tokens: everything you send (prompt, files, conversation history).
Cached input tokens: eligible reused prompt segments are billed cheaper (when prompt caching hits).
Output tokens: everything the model generates (code, prose, tool traces, etc.).

Two practical implications:

Output is often the biggest driver for agentic runs that generate lots of code.
Caching is the biggest lever for long, multi-turn sessions where the “static” part of your prompt stays unchanged.

Pricing Snapshot (May 5, 2026)

All prices are USD per 1M text tokens.

Codex Models (API)

Model	Standard input	Cached input	Standard output
`gpt-5-codex`	$1.25	$0.125	$10.00
`gpt-5.1-codex`	$1.25	$0.125	$10.00
`gpt-5.1-codex-max`	$1.25	$0.125	$10.00
`gpt-5.2-codex`	$1.75	$0.175	$14.00
`gpt-5.3-codex`	$1.75	$0.175	$14.00

OpenAI also lists Priority pricing for gpt-5.3-codex at 2x standard (input $3.50, cached input $0.35, output $28.00 per 1M tokens).

Flagship GPT Models Commonly Used for Coding (API)

Model	Standard input	Cached input	Standard output	Context window
`gpt-5.2`	$1.75	$0.175	$14.00	400k
`gpt-5.4`	$2.50	$0.25	$15.00	1.05M
`gpt-5.5`	$5.00	$0.50	$30.00	1M
`gpt-5.5-pro`	$30.00	-	$180.00	1.05M

Sources: the model cards for gpt-5-codex, gpt-5.1-codex(-max), gpt-5.2-codex, and gpt-5.3-codex, plus the official API pricing table.

Note: gpt-5.4 has special long-context pricing rules for very large prompts; check the model page before budgeting a multi-hundred-K token single request.

Context Windows: Why the Numbers Don’t Always Match

If you compare API docs to what you see inside ChatGPT or Codex, the context windows can look inconsistent. That’s expected:

API context windows are defined per model ID (what you can send in a single request).
ChatGPT context windows depend on the surface and sometimes the plan tier (Instant vs Thinking, plus plan limits).
Codex can expose a different context window than the API for the same family (for example, GPT-5.5 in Codex is listed with a 400K context window).

When you’re writing a budget, always anchor on the surface you’re paying for: API token billing vs ChatGPT subscription usage.

Capability Timeline (What You Were Paying For)

A simplified view of how the coding-focused lineup progressed:

Sep 15, 2025: GPT-5-Codex introduced as a GPT-5 variant optimized for agentic coding in Codex.
Sep 23, 2025: GPT-5-Codex became available via API key in the Responses API.
Nov 19, 2025: GPT-5.1-Codex-Max released for longer, more detailed, long-running tasks (with compaction-style long-horizon work).
Dec 18, 2025: GPT-5.2-Codex released with stronger long-horizon performance and significantly stronger defensive cybersecurity capabilities.
Feb 5, 2026: GPT-5.3-Codex released as the most capable Codex model to date, with faster operation and improved frontier agentic performance.
Mar 5, 2026: GPT-5.4 released across ChatGPT, API, and Codex as a frontier model for coding and professional workflows.
Apr 23-24, 2026: GPT-5.5 introduced and rolled out broadly (ChatGPT, Codex), with API availability shortly after.

Capabilities vs Price (A Practical Matrix)

All the models below can be used for coding, but they trade off capability, context, and cost differently.

Model	Surface	Context window	Max output	What you’re buying
`gpt-5-codex`	API	400k	128k	Baseline Codex behavior at the lowest Codex pricing.
`gpt-5.1-codex-max`	API	400k	128k	Best value when you want long-running, detailed coding work at older Codex pricing.
`gpt-5.3-codex`	API	400k	128k	Highest Codex capability at mid-tier token pricing.
`gpt-5.2`	API	400k	128k	Previous frontier model pricing, often “good enough” for most coding tasks.
`gpt-5.4`	API	1.05M	128k	Much larger context; better when your repo/tooling context is huge.
`gpt-5.5`	API	1M	128k	Higher cost, but aimed at better token-efficiency and fewer iterations on complex work.

The Cost Story: What Changed From Launch to Today

From GPT-5-Codex to GPT-5.3-Codex, the headline change is a 40% price increase for standard token rates:

Input: $1.25 -> $1.75 per 1M tokens
Output: $10.00 -> $14.00 per 1M tokens

So your decision is less about pennies and more about whether the newer models let you:

finish tasks in fewer iterations
generate fewer wasted tokens (less thrash)
avoid human time (review cycles, debugging time, incident risk)

OpenAI claims GPT-5.3-Codex uses fewer tokens than prior models on key agentic benchmarks, which matters because token efficiency can offset higher per-token prices.

For flagship GPT models, the jump is larger:

gpt-5.2 -> gpt-5.5: input goes from $1.75 -> $5.00 per 1M tokens, and output from $14.00 -> $30.00.

The bet you’re making with gpt-5.5 is that you get fewer retries and less wasted output, so the effective cost per merged PR (or resolved ticket) drops even if per-token rates are higher.

Practical Levers to Reduce Token Spend

If you’re cost-sensitive, optimize the shape of the conversation, not just the model:

Constrain output: ask for a git diff (or patch) first, and keep explanations short until after the patch is correct.
Avoid re-sending the world: keep stable instructions and context fixed so prompt caching has a better chance to hit.
Split tasks: run separate, smaller prompts for “analysis” and “implementation” instead of one massive all-in-one run that bloats output.
Measure output: agentic workflows can accidentally print long logs; treat “verbosity” as a cost setting.

Quick Cost Examples (Standard Pricing)

A simple estimate:

cost = input_MTok * input_price
     + cached_input_MTok * cached_input_price
     + output_MTok * output_price

Example A: “PR review + patch” sized run

Input: 200k tokens (0.20 MTok)
Output: 20k tokens (0.02 MTok)
No caching

Estimated cost:

gpt-5-codex: 0.201.25 + 0.0210.00 = $0.45
gpt-5.3-codex: 0.201.75 + 0.0214.00 = $0.63

Example B: long multi-turn session with caching

Input (cached): 800k tokens (0.80 MTok) reused across turns
Input (new): 200k tokens (0.20 MTok)
Output: 50k tokens (0.05 MTok)

Estimated cost:

gpt-5-codex: 0.201.25 + 0.800.125 + 0.05*10.00 = $0.85
gpt-5.3-codex: 0.201.75 + 0.800.175 + 0.05*14.00 = $1.19

Caching keeps these sessions sane. Without caching, Example B would be materially more expensive.

Which Codex Model Should You Choose (Cost-First)?

As of May 5, 2026:

Default to gpt-5.3-codex when task success probability matters (large refactors, tricky bug hunts, security-sensitive changes).
Use gpt-5-codex / gpt-5.1-codex when you want Codex behavior at lower cost and can tolerate more iterations.
Reserve Priority for interactive workflows where latency is worth the 2x token rates.

If you track one metric, track cost per merged PR (or cost per resolved ticket), not cost per request.

ChatGPT Plus Comparison (May 2026)

Token pricing applies to the API. ChatGPT Plus is a subscription with usage limits that can change over time, but it matters because many teams do “explore in ChatGPT” and “ship via API”.

What Plus is typically used for:

Interactive debugging, brainstorming, and quick code generation
Short agentic workflows where you don’t need per-request cost accounting

What to keep in mind:

Plus includes access to newer frontier models in ChatGPT (for example, GPT-5.4 Thinking and GPT-5.5 Thinking rollout notes mention Plus availability).
ChatGPT context windows and limits differ by mode (Instant vs Thinking) and tier; don’t assume they’re the same as the API context windows.

A practical comparison for ChatGPT Plus (subject to ongoing product changes):

Mode in ChatGPT	Typical Plus context window	Notes
GPT-5.3 Instant	32K	Fast, general chat/coding help.
GPT-5.5 Thinking	256K	Larger context; Thinking mode has its own limits and is different from API context windows.

If you need strict cost control and repeatability, prefer the API. If you need convenience and don’t want to meter usage per request, Plus can be a good default for day-to-day coding assistance.